OpenAI's Pando Cuts Git Push Latency 80%

OpenAI has moved all development in its internal monorepo onto Pando, a Git-compatible server backend it built in 10 weeks with its own Codex coding agents, the company said in a Nov. 2 engineering post titled "Scaling Git at OpenAI." OpenAI said the rebuild cut median push latency by nearly 80% as agent-driven pushes overwhelmed its previous Git hosting.

Key takeaways

• Pando serves nearly 1 petabyte of Git data a day and began handling production traffic in April 2026, according to OpenAI.
• OpenAI said reliability of its old Git setup at times fell below 90% as Codex agents multiplied branches and pushes.
• The backend stays internal, but OpenAI's public Git fork tracks 50 patches prepared for upstream Git.

Why OpenAI rebuilt its Git backend

Coding agents create far more branches, pushes and clones than human teams. OpenAI said that starting around the end of 2025 its monorepo suffered slow fetches, slow pushes and failed merges, with reliability at times below 90%.

Rather than switch version-control systems, OpenAI kept the Git protocol. Engineers and agents still use the standard Git client, and Pando replaces only the server side.

How Pando works

Pando is written in Rust. Each push is stored as an immutable packfile in a blob store, while a metadata service tracks a version for every branch. Background workers merge small packs into larger ones, a design similar to a log-structured merge tree.

The key change is ordering. Instead of a single repository-wide sequence, each branch updates independently under what OpenAI calls causal consistency, so pushes to unrelated branches no longer wait on each other. Fetches use Git's packfile-URI feature to hand clients download links for stored packs.

Pando performance figures

OpenAI reported these production numbers, which have not been independently verified:

• Median push latency down nearly 80%, or almost five times faster.
• Nearly 1 petabyte of Git data served per day "with just a handful of CPU cores."
• A peak of more than 57 terabytes of packs in one hour, with over 99% of bytes served from local disk cache.
• Clone and fetch speeds that can saturate 10-gigabit links on standard cloud virtual machines.

OpenAI said agents helped check the design. Formal Lean and TLA+ models were written by agents using GPT-6 Astra, alongside conformance tests and deterministic simulations.

Context: AI agents strain code hosting

OpenAI is not alone. Anthropic said on Oct. 13 that it had moved its monorepo from Git to an in-house system called Scion after daily merges topped 30,000 in September. GitHub's Octoverse report said agents started more than 40% of pull requests on the platform in September, up from 17% in March.

The two companies took different routes. Anthropic left Git, while OpenAI kept Git clients and rebuilt the server. OpenAI's fork at github.com/openai/git carries 92 commits on top of upstream, and its patch tracker listed 50 patches for upstream as of Oct. 31. Some, including packfile-URI download fixes and sparse-index support for git stash, are already in upstream master.

What's next

OpenAI said it is "exploring ways to share the technology" but gave no release date or pricing. It named large-scale collaboration among many agents on one shared repository as the next workload Pando must handle.

FAQ

Is OpenAI's Pando open source?

No. Pando is internal. OpenAI said it is exploring ways to share it, and it is sending Git client fixes upstream through its public fork.

Do developers need a new Git client to use Pando?

No. Pando keeps the Git protocol, so OpenAI engineers and agents use standard Git commands and clients.

How long did Pando take to build?

OpenAI said it took 10 weeks from kickoff to production, with one engineer designing the core and agent teams implementing it.

References

• OpenAI Git · patches & notes — OpenAI (as of Oct. 31, 2026)
• openai/git — GitHub
• Pando — OpenAI rebuilt its Git backend — AI/TLDR (Nov. 2, 2026)
• Scaling Git at OpenAI: Pando Explained — Kingy AI